Name versus engine
DeepNude AI: same name, different engines underneath
Calling something AI says nothing about what runs underneath. What differs most between these services is how much they are willing to put in writing.
Short answer
No model sees through fabric. What happens technically is different: part of the photo is masked, its contents are thrown away, and the empty area is drawn from scratch by a generative model. The body in the result was not recovered from the original picture, because that information was never in the file — the shape is invented from patterns in millions of other images.
This page takes apart the engine behind the services that put "AI" next to the name: how diffusion and inpainting work, where the output falls apart, and which claims cannot be checked from outside at all. The wider map of what is left of the original program and what replaced it sits on the front page.
The engine
Diffusion and inpainting, not an X-ray
Two separate techniques work together. Diffusion does the drawing; inpainting decides where the new drawing is allowed to appear. Without the second one, a result like this would never sit flush against the source photo.
A diffusion model is trained in a way that sounds backwards. It is shown a complete image, that image is degraded with noise step by step until it is random speckle, and the model learns to walk it back. Once trained, the direction is reversed for production: the model starts from pure noise and cleans it a little at a time until something plausible stands there. Not one of those steps reads hidden information out of the photo — what it reads is the visible pixels around the masked area.
What makes it cheap and fast is where the arithmetic happens. The latent diffusion paper by Rombach and colleagues (2022) moved the process out of pixel space and into the latent space of a pretrained autoencoder, and the same paper reports a new state of the art for image inpainting. Those two sentences explain the whole industry now standing on top of it: without that compression, one image would need hardware that does not fit inside a business model charging cents per picture.
The mask is drawn
The service marks the area to be replaced, usually automatically through body detection. Everything outside it is locked and left untouched.
The contents are discarded
Inside the mask nothing survives from the original photo. The starting point for what follows is noise, not thinned-out fabric.
The noise is cleaned in steps
The model guesses, then corrects its guess a few dozen times. Each step is steered by the pixels outside the mask and by a text prompt the service wrote for you.
The seam is blended
The mask border is dissolved, skin tone and light direction are matched to the surroundings. That cosmetic step is what makes the result read as one photograph.
The image is decoded
The output is decoded back to full size. That size is chosen by the service, and almost none of them writes it down before you pay.
Which is why the vocabulary used in the advertising is wrong from the first word. Nothing is uncovered and nothing is revealed: part of an image is deleted and refilled by a general-purpose drawing engine.
Failure points
Where these models fall over
The failures have a pattern, and the pattern comes straight from the mechanism above. The model is strong on large smooth surfaces and weak on anything with structure that has to stay consistent.
| Where the mask meets the real photo | What shows up in the result |
|---|---|
| Objects crossing the body | A sling bag, a backpack strap, long hair, folded arms: all of them land on the mask border, and there the model has to guess what is in front and what is behind. |
| Patterned fabric and folds | Stripes, checks, embroidery, a draped terno sleeve. The pattern has to be cut exactly at the mask edge, and that cut is almost always visible when the image is enlarged. |
| Tattoos, moles, scars | Present in the source, gone from the result, or moved somewhere else. The model keeps no record of any particular person's skin. |
| Fingers and jewellery | The classic weak spot of image models: the count of fingers changes, a ring fuses into skin, a bracelet snaps in the middle. |
| Angled poses and cropped bodies | The less context there is around the mask, the freer the invention, and the further the proportions drift. |
| Low-resolution sources | The output cannot be sharper than the input. What grows is the pixel count, not the detail. |
| More than one person in frame | The mask spreads to the person standing next to them, and body parts migrate between figures. |
| Hard light and sharp shadows | Light direction in the new area is recomputed, so shadows on the body often disagree with the shadows on the face. |
That list has a direct effect on cost. Because the failure is only visible once the image is finished, one photo rarely takes one run, and not one of the six services examined here offers a free retry. The real arithmetic is not the price of one image but the price of one image multiplied by the number of attempts until you stop. How those units come off the balance, and how a service that rents someone else's model bills for it, is worked through on the page explaining that generating is not editing.
Six things worth understanding
What changed and what stayed the same
The technology has turned over twice since , but the questions that decide anything are the same: who computes, where, at what price, and what it means in front of Philippine law.
From GAN to diffusion
The program belonged to a different family of models. According to the Motherboard report that first tested it, it was built on pix2pix, an open-source generative adversarial network, and trained on more than 10,000 photographs for one narrow task. The consequence showed in the output: a model like that only works on the kind of frame its training data contained, and breaks immediately outside it.
Diffusion changed the economics. The base model is general-purpose, built by someone else, and steered towards a narrow task with small adjustments and a text prompt. What the services sell today is not a model but an interface, a queue and a billing system sitting on top of another company's model. That is also why dozens of services could appear inside a year when in 2019 there was one.
Which model each one claims
Asked directly, five of the six say nothing. Not one document published by Ainudez, Undresswith, UndressHer or Deep Undress names an engine, a version or an owner, and Pornworks publishes nothing readable in any direction. Adjectives are the whole of it — advanced, powerful, ours — which means there is no claim on the table that could be checked, confirmed or caught out.
The exception is the service that processes no photographs of people. WaveSpeed sells access to more than a thousand image, video and audio models built by other companies — ByteDance, Alibaba, Google, OpenAI and MiniMax among them — with the catalogue and the per-execution rate readable before an account exists, and with terms describing those models as third-party offerings it is not a party to. It is the only one of the six where the answer to "whose model is running" is printed rather than implied.
Read those two paragraphs together and the pattern is the wrong way round: the service that admits it built nothing is the only one that identifies what it is selling, while the services presenting the engine as their own name nothing at all. The program, for its part, said openly that it was built on pix2pix. Seven years of progress have made the engines far better and the disclosure considerably worse.
Where the shape comes from
What appears inside the mask is a weighted average of training data, not a reconstruction of the person in the photo. That is why the output often looks generic: proportions drift towards the dataset mean, skin comes out too even, and whatever is distinctive about that person is gone. The further a real body sits from that average, the more obviously the result misses.
The reverse question matters just as much: does the image you send become training data. Here the six do not agree. Ainudez states in its privacy policy that inputs and outputs are not used for marketing, profiling or third-party model training. WaveSpeed does the opposite and lists uploaded material as training data, acknowledging it may contain sensitive information. UndressHer says its training uses synthetic data rather than user images. All three are self-declarations with no outside audit.
Who does the computing
Diffusion needs a graphics card, and that closes one question straight away: no service under this label runs inside a phone. The photo always leaves, is processed on somebody else's server, and comes back. Whatever is installed on your device, if anything, is a window.
So the difference between opening a browser tab and installing an app is not about where the processing happens — it is about how much device access you hand over. The other half of the comparison, tariff by tariff, is the page on price and deletion windows.
Why footage costs so much more
Video is a run of images that have to agree with one another, so the cost multiplies rather than creeps. The Ainudez tariff shows it in numbers: one still is 15 credits, while seven seconds at 720p is 300 credits — twenty times as much, for seven seconds. Above 720p its price list offers nothing at all.
Frame-to-frame consistency also introduces a new failure mode: flicker, trembling edges, and shapes that change halfway through a movement. What each service promises about clips, and what it actually writes down, is separated out on its own page.
Claims that cannot be tested
Three promises show up on the front page of services like this, and none of the three can be checked from outside: realistic results, photos are never kept, and the best model. The first has no unit. The second can only be believed or not, because nobody can look inside another company's servers. The third usually points at the same third-party model a competitor is calling.
What can be checked is the boring material: whether the tariff is readable before sign-up, whether a legal entity is named, whether there is a deletion deadline in days, and whether a payment can be disputed. Those questions are also what decides where billing starts once the opening balance is gone.
Invented images are still punished
The fact that the body in the result is fiction does not improve anyone's position under Philippine law. Section 12 of the Safe Spaces Act of 2019 lists, as gender-based online sexual harassment, the uploading and sharing without the victim's consent of any media containing photos, voice or video with sexual content, and any unauthorised sharing of a person's photos or information online. Section 14 sets the penalty at prisión correccional in its medium period or a fine of ₱100,000 to ₱500,000, or both.
What is judged is the conduct towards that person, not the technical quality of the file. A result that obviously looks fake still attaches somebody's face to sexual material and still circulates under their name. "How realistic is it" is therefore not a legal question at all, and answering it proudly mitigates nothing. The money side of the same decision — subscriptions, non-reversible payments and the charges that will not cancel — is collected separately.
Three that publish numbers
The services that show the figures before you go in
All three run in a browser and publish a tariff or a cost structure outside the account. Everything below is quoted from their own documentation, not from testing on our side.
-
1Model catalogueWaveSpeed
Rents out image and video models built by other companies, billed per execution, with unit prices readable before registration — according to its own documentation.
-
2Photo and videoAinudez
Sets 15 credits per image and 50 to 300 credits per short clip, and states that user files are not used to train third-party models — according to its own documentation.
-
3Photo and videoDeep Undress
Accepts JPG, PNG, WEBP and BMP up to 20 MB and video under 50 MB, and sells access as a one-off purchase with no auto-renewal — according to its own documentation.
The "Open site" buttons are affiliate links: if you buy something on the other side, this desk earns a commission at no extra cost to you, and it changes nothing on this page. We processed no images and opened no paid accounts, so every line on a card is a quotation from a service's own documents. One thing none of the three publishes, and the other three do not either: the output resolution for a still image.
Common questions
What people ask about the engine
The answers below rest on two things only: technical publications anyone can open, and the documents each service publishes about itself.
Does the model actually see the body under the clothing?
No, and physically it could not. A photograph records only light reflected off the outermost surface. The model deletes the masked area and redraws it from noise, so what comes out is an invention consistent with its surroundings, not information hidden inside the file.
Why do two runs on the same photo give different results?
Because the process starts from random noise. A different starting point produces a different cleaning path. There is no single answer the model is looking for, which is also why improving a result always means paying again.
Can the output be told apart from a real photograph?
Often yes, by enlarging the mask edge: a cut-off fabric pattern, the wrong number of fingers, shadows pointing in disagreeing directions, a missing tattoo. But those signs thin out at low resolution, and on a phone screen many of them are invisible.
Are the models retrained on images users send?
It depends on the service, and the answer is only as strong as the document. One states plainly that it does not use inputs to train third-party models; another lists uploaded material as training data; a third says its training uses synthetic data. None of the three can be verified from outside.
Could any of this run offline, on the phone itself?
Not in the form that circulates now. A diffusion model needs video memory and storage far beyond the size of an ordinary app. Anything claiming to work on-device still sends the photo to a server, and the size of its own installer gives that away.
Does raising the resolution make the result more accurate?
Bigger is not the same as more accurate. Upscaling adds pixels, not correct detail, and inside the masked area there is no original detail to restore. What usually grows is the artefacts along the seam.
Did this desk test the output of these services?
No. We sent no images and produced no results. What we read and archived are the services' public documents plus technical publications, and where they say nothing, this page records the gap instead of filling it with a guess. The origin of the name itself, and why it lasted four days, is covered separately, as is the question of whether the labels and the process behind them have anything to do with each other.